Meta AI slaps a second agent on failing bots — and scores jump 10 points

AI agents have a nasty habit of going senile mid-task. They’ll flag a constraint, then break it two steps later while fixing something else. They’ll diagnose an error pattern, then rediscover it like it’s brand new. Meta AI researchers have a name for this: behavioral state decay. It’s not that the information vanishes — it’s buried in the context window, drowned out by noise. And no, they found, just giving agents longer memories doesn’t fix it.

So Meta built a second AI that acts as a dedicated memory coach. The system pairs a standard action agent with a separate memory agent that watches a sliding window of recent steps and maintains a structured memory bank. Crucially, this coach isn’t a backseat driver — it only chimes in with targeted reminders when it decides the agent is about to repeat a mistake. It can also choose to say nothing at all. “Choosing not to intervene is part of the policy,” the researchers note, and that silence option turned out to be essential.

The results are tangible. When paired with Claude Sonnet 4.5 as the action agent, the system solved 46% of Terminal-Bench tasks on the first try, up from a 38% baseline. On tau2-Bench, which tests conversational tool use across airline, retail, and telecom domains, the task-weighted average climbed from 55% to 62%. The gains weren’t uniform — airline and retail jumped roughly 10 points each, while telecom barely budged — suggesting the memory agent intervenes more aggressively when tasks have clearer procedural pitfalls.

What’s most interesting is that the system still helped a stronger model. Claude Opus 4.6 improved by 2.4 points on Terminal-Bench and 2.5 on tau2-Bench. This isn’t just a crutch for weak models. When the team stripped out the silence option and force-fed the action agent reminders at every step, performance actually dropped. Constant nagging creates its own noise problem. The selective approach — knowing when to shut up — is the real innovation here, and it points toward a future where agent architectures aren’t just about bigger models but smarter coordination between them.